Following the example below and looking the nms source code, I expected a NaN error, as the intersection and union will be zero.
import torchvision # torchvision==0.5.0+cpu
import torch # torch==1.4.0+cpu
boxes = [[0.0, 0.0, 1.0, 1.0],
[2.0, 1.0, 1.0, 2.0]]
boxes = torch.tensor(boxes)
scores = torch.tensor([1., 0.5])
keep = torchvision.ops.nms(boxes, scores, 0.7)
If this same example is used with this nms implementation (removing the +1 from the source code to be equivalent to the torchvision implementation), it raises a NaN error as expected.
Am I missing something ?
Thanks.